Explainable post-training bias mitigation with distribution-based fairness metrics